High-precision apple classification and traceability based on enhanced CBAM for near-infrared spectroscopy

Area of Science:

• Agricultural Science
• Food Science
• Computer Science

Background:

• Apple origin traceability is vital for food safety and consumer trust.
• Traditional spectral analysis methods struggle with accuracy due to limited feature extraction and class imbalance.

Purpose of the Study:

• To develop an accurate and robust deep learning model for apple origin classification.
• To enhance feature extraction and address class imbalance in spectral data analysis.

Main Methods:

• Collected 2400 near-infrared spectra from 200 apple samples across four varieties.
• Applied Multiplicative Scatter Correction (MSC) for spectral data preprocessing.
• Developed a 1D-CNN model with an enhanced Convolutional Block Attention Module (CBAM), multi-scale convolution, dense residual connections, and Balance Softmax Cross-Entropy loss.

Main Results:

• Achieved a high accuracy of 97.12% ± 0.74% in apple origin classification.
• The enhanced CBAM module with triple pooling and multi-scale convolution improved spectral feature selection.
• The Balance Softmax Cross-Entropy loss effectively handled class imbalance.

Conclusions:

• The proposed 1D-CNN model with CBAM significantly improves apple origin traceability.
• This approach offers a robust solution for food safety and agricultural product authentication.
• Advanced deep learning techniques can overcome limitations in traditional spectral analysis for agricultural applications.